GWAS and TWAS are two main ways to find underlying drivers of phenotypes. GWAS looks for associations between genetics (G) and phenotype (Y) while TWAS looks for associations between gene expression (E) and phenotype.
The idea behind TWAS is that genetics control expression and then expression controls phenotype and so by regression phenotype on expression, on is aggregating the genetic effects to make a more powerful study.
However, it is not guaranteed that a TWAS hit implies that the expression level is driving the phenotype, it may just be that the expression level is tagging coding variation that is driving the phenotype.
However, this is testable. If expression level is driving trait, then the effect of every expression quantitative trait loci (eQTL, a genetic variant that effects gene expression), should have proportionally the same effect on phenotype. If they do not, then it is not expression driving the phenotype.
Y ~ G*beta*gamma
where beta is the effect of genetics on expression and gamma the effect of expression on phenotype.
Therefore, for each known eQTL in a TWAS gene, the beta is known and the estimated gammas should all be the same. Deviation from this would imply that expression is only tagging the underlying causal variation.